Over the past year, many SaaS products have added AI chatbots. They answer questions, guide users, and reduce some support load. That was a useful first step, but it is no longer enough.

Most AI still live in a chat window with no awareness of the application. They don't understand product state, selected data, user permissions, or the current workflow, so users end up restating context the system already has.

That approach does not scale well for real products.

We’ve always believed copilots are the way to deliver the best customer experience—not by answering questions, but by actually doing things

We are releasing Copilot SDK as an open-source toolkit to explore this idea and make context-aware, action-driven copilots for product teams.

What it does:

- Understands application state including current page, selected data, and user permissions

- Executes backend and frontend functions instead of only responding with text

- Delivers richer product experiences through generative UI such as tables, forms, and interactive buttons

- Understands user workflow and intent based on in-product context

- Maintains session context so interactions remain consistent

Example: Instead of the AI asking "What do you need help with?", it understands the user context is viewing failed transactions from last week and can immediately offer to retry them, export the data, or investigate patterns.

Technical details:

- Works with React, Next.js, Vite (Vue & Angular coming soon)

- LLM-agnostic (bring your own model)

- State injection via context providers

- Tool execution layer for safe function calling

- Full data ownership (everything runs in your infrastructure)

Docs + examples: https://copilot-sdk.yourgpt.ai

Happy to answer technical questions about implementation, or specific use cases.